The gist: instead of shipping an installer for an app with a fixed codebase - same binary, same experience, same constraints for everyone - you ship a description of what the software should do. An AI agent reads that set of suggestions and builds a custom app based on your situation, namely your goals, infrastructure & constraints.
You can think of it like an architect shipping a conceptual blueprint for a house with aspects that are open to interpretation by the builder. The builder comes briefed on the buyer’s wants & needs and puts his spin on the architect’s original plan to create a bespoke house for the client.
The old model of parameterizing settings of a cloned codebase
Every piece of software you’ve ever installed was a clone. Same installer, same binary, same app. Maybe you tweak some config settings, pick a theme, toggle a few preferences. But underneath, it’s the same code base running on every machine and simply adjusting parameters to yield a slightly-modified variant of the experience to the user.
That made sense when humans had to write every line of code. Write once, ship it to millions. Economies of scale.
The equation changes here though when the cost of writing custom software approaches zero with agentic coding platforms.
A sufficiently well thought-out idea file is the ultimate installer
Karpathy’s proposal: what if instead of shipping the building, you ship a blueprint that’s open to interpretation by the consumer? Not even a detailed blueprint - more like an architectural philosophy. “Here’s what this system should do, here’s how it should be structured, here are the constraints.”
Then an AI agent - Claude Code, Codex, whatever you’re using - reads that spec and builds it fresh and builds it just for you. It factors in your existing tools, your preferences, your technical context, your actual needs.
Software 2.0 looks more like sexual reproduction
Maybe a better analogy for what Karpathy is proposing is not blueprints and buildings but more about cloning vs sexual reproduction from two parents.
The software provider contributes one set of chromosomes - the specification. The goals, the architecture, the patterns, the constraints.
The consumer contributes the other set - their unique context. Their existing infrastructure. Their preferences. Their specific problem. Their resources.
What comes out is a software baby conceived uniquely from the ground up that carries the DNA of both parents but is identical to neither. And just like actual offspring, it’s a better fit for its environment than either parent would produce alone.
Every “install” is non-deterministic and produces something different that’s uniquely yours and adapted to your scenario.
We’re already doing this at Vibecode Lisboa
We’ve been doing a version of this at Vibecode Lisboa with the Confabulator tool.
When a student goes through the Confabulator, they describe their project in plain English. The tool takes that description - their chromosomes - and combines it with our template architecture - our chromosomes - and produces a unique GitHub repo. Same idea file. Different user DNA. Different software baby every time.
Ben, a non-technical student in Build School Cohort 2, used the Confabulator and had a deployed micro-SaaS for investors in two days. His app looked nothing like the app I built the same weekend (Prep.Training) using Confabulator. Same tool, same birthing process, completely different output because our unique inputs shaped the result.
Why this matters for you
If you’re building software: think about whether your distribution model should shift. Instead of shipping apps, maybe you ship idea files. Let the user’s agent build what they actually need.
If you’re a non-coder: this is the future getting closer to you. The gap between “install this app that kind of does what I want so I’ll accept the tradeoffs” and “describe exactly what you need and an agent builds it” is closing fast.
If you’re somewhere in between: this is why learning to work with AI agents matters. The person who can clearly articulate what they need and mesh that with a promising idea file is the person who can generate custom software that fits like a glove. The person who can’t is still tracking down installers and settling for shoe-horning his/her business into a shoe that knowingly doesn’t fit right.
Where do we go from here
We are living in the future. The software you use tomorrow might share DNA with a million other installations but be genetically unique to you. Karpathy sees it. Dorsey sees it. The 16 million people who viewed his tweet see it.
If you’re shipping software consider distilling the essence of your application to an idea file that your consumers can give to their agent to birth the perfect app for their needs.
And if you’re a non-technical user seeking to wade into these waters and test this stuff out, take five minutes and create your first Confabulator project and experience this new “software birthing” process first-hand.
Excellent read, thanks Sean! :)